메뉴
Fetch the latest PR review comments for the current branch, analyze and summarize feedback, and help the developer plan fixes. Use when the user mentions PR comments, review feedback, reviewer requests, or wants to understand and address PR review comments.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
| name | gh-pr-comment-assistant |
| description | Fetch the latest PR review comments for the current branch, analyze and summarize feedback, and help the developer plan fixes. Use when the user mentions PR comments, review feedback, reviewer requests, or wants to understand and address PR review comments. |
Analyze PR review comments for the current branch, summarize feedback, and help the developer plan fixes or apply suggestions if necessary.
BRANCH=$(git branch --show-current)
PR_NUMBER=$(gh pr view --json number -q '.number')
gh pr view --json number,url,title,baseRefName
If no PR exists for the current branch: Inform the user and suggest gh pr create. Stop. If gh is not
authenticated: Show error and suggest gh auth login. Stop.
Fetch both inline review comments and general PR comments:
# Get repo info
REPO=$(gh repo view --json nameWithOwner -q '.nameWithOwner')
# Inline review comments (code-level)
gh api repos/${REPO}/pulls/${PR_NUMBER}/comments
# General PR comments and review summaries
gh pr view --json comments,reviews
Group comments by file and thread. Tag each as:
Skip already-resolved threads.
For each comment or thread:
Clarification gate: If a comment is ambiguous, use the Ask User Questions tool to ask the user before proposing a resolution.
Present all comments using this template:
### Comment #N: [file:line] by @reviewer
**Type:** Blocking | Suggestion
**What they asked:** <1-sentence summary>
**Proposed action:** <code change | decision | no-op> -- <brief reasoning>
After presenting the analysis, switch to Plan mode and use the Ask User Questions tool with one question per comment group. Each question should:
Options should be specific to the comment, not generic labels. For example, a comment about error handling might have: "Add try-catch to the fetch call", "Discuss whether we need error boundaries here", "Investigate what errors are possible", "Skip".
Work through comments based on the developer's direction. For each comment being addressed:
Confirmation gate: Wait for explicit approval before applying any code changes.
Repeat until the developer is satisfied or decides to continue on their own.
Audits and fully configures how a website looks when shared (Open Graph, Twitter/X cards, WhatsApp/Discord/Slack previews) and in the browser (title, favicons, theme-color, web manifest), generating missing assets and verifying the result locally. Use when setting up link previews, og:image, favicons, social sharing metadata, or when a shared link "looks wrong" on a platform.
Iterative React-health loop: inventory every React app in the repo, baseline with the latest react-doctor (pinned for the rest of the run), fix one root cause per iteration, and re-verify until every app scores 100/100 or every remaining finding has a user-approved written justification. Never silences findings with ignores, exclusions, or deleted behavior, and stops itself after two iterations without measurable progress. Use when the user mentions react-doctor, React health score, React anti-patterns cleanup, or driving a React codebase to a clean scan.
Iterative dependency-CVE remediation loop: scan with ecosystem-native advisory tooling, assess whether each high/critical finding is actually reachable (with call-path evidence, not vibes), fix the highest-risk reachable one with the smallest credible change, re-verify, and repeat. Terminates when no exploitable high/critical CVE remains or every remaining finding has an evidence-backed reachability assessment and an approved risk decision. Use when the user mentions CVEs, vulnerability scanning, dependency security, npm/pip/cargo audit, security patching, or wants a recurring dependency-security loop.
Orchestrates a comprehensive frontend UI review by running parallel analysis across typography, layout, accessibility, responsiveness, copy, visual polish, and design critique. Uses browser inspection for visual context. Produces a prioritized plan of small, concrete fixes -- never rewrites. Use when the user mentions UI review, frontend review, page review, design review, visual review, audit my page, or review my UI.
Review open PRs authored by the user for a given repo, showing title, status, review state, open comments, CI checks, and recommended next steps. Use when the user mentions PR review, my PRs, open PRs, PR status, PR dashboard, or wants an overview of their pull requests.
Surfaces the single most important blindspot in a plan, decision, or situation by checking whether the frame is right, generating candidates across 8 lenses (hidden constraints, false assumptions, second-order effects, stakeholders, blast radius, base rates, time horizon, counter-positions), then either committing to one or returning a calibrated "no blindspot" when none is load-bearing -- never a list of risks. Use when the user asks "what am I missing", "what's the most important thing I'm missing about this situation or decision", "what's my blindspot", "what's the biggest risk I'm not seeing", or wants a single sharp second opinion before committing.